You are on page 1of 6

The Rapid Method for Road Extraction from High-Resolution Satellite Images Based on USM Algorithm

Liu Xu, Tao Jun, Yu Xiang, Cheng JianJie, and Guo LiQian Email: Lxfeng_2005@163.com AbstractWith the rapid development of 3D-GIS and image extraction technology, the automatic updates of GIS information have become possible. The road information is an important part of the GIS data, and the general road extraction algorithm had the problem of computational complexity and was time-consuming. In this paper, the improved Unsharp Mask (USM) sharpening algorithm is used to enhance the image, and to strengthen the road feature with the surrounding environment color gradient. A rapid method of road extraction from satellite images based on edge detection was achieved by the operations of image preprocessing, threshold segmentation, corrosion and expansion. The algorithm was achieved by matlab and applied to high-resolution satellite images. The experiment results proved that this method could effectively reduce the noise of non-target and achieve accurate road extraction from the general images. KeywordsUnsharp Mask (USM); road extraction; edge detection; high-resolution satellite image

Northwest Institute of Nuclear Technology, Xian, 710024, Shanxi, China

I.

INTRODUCTION

Image processing and remote sensing technology of the long-term pursuit is an important goal, and it is to realize the spatial information acquisition of automation and intelligent processing. The artificial surface features are the spatial geographic information database of important elements. The artificial features, mainly including buildings, roads, and bridges, etc. Identification and extraction of buildings and roads accounted for 80% of the high-resolution satellite images. Buildings and roads have obvious locating features, and identification and accurate positioning of them for feature extraction, feature matching, image understanding, mapping, and other objectives of the reference body has important significance. According to the degree of automation of extracting linear features (mainly this refers to the road), road extraction is divided into two methods automatic and semi-automatic. Undoubtedly, fully automatic extraction of remote sensing image target recognition and extraction of the direction of development is the ultimate goal. At present, many researchers from various aspects put forward various research methods, resulting in a number of practical results. Now, forming methods, Wang Zhen-Guo based his research on the distance constraints of high resolution satellite road

extraction method [1]; Wang Mins research is based on scan the snake, a semi-automatic method of extracting road from high spatial resolution remotely sensed imagery [2]. Ye Qin proposed a method of road extraction based on regional growth and the shape of the space constraints [3]. Xia Jing et al. proposed a method of road extraction based on the Hop-Along algorithm from high resolution remote sensing images. Through matlab experiment of high-resolution remote sensing images, they used the partial Hop-Along algorithm for multi-line fitting to get a single pixel of the central line of the road, realizing road extraction [4]. In the high-resolution satellite images, the road has more performance for the strip surface domain feature, not simply as a linear target processing. At the same time, high-resolution satellite images contain relatively few spectral information and relying on spectral information is not very good for road extraction. Due to the improvement of image resolution, the image has all sorts of detailed information to obtain very good reflection, but these details characteristics for the extraction of the road takes more time and is a kind of interference. Also, high-resolution satellite images have shadows that are more serious, particularly in urban areas, which have an impact on road extraction. General road extraction, mainly from image extraction method, needs to improve its extraction algorithm, and in this paper, based on a high-resolution satellite image processing approach, proposes rapid technology for road extraction based on USM sharpening algorithm. The basic idea is: use the Unsharp Mask (USM) sharpening algorithm to enhance the image, and strengthen road and surrounding environment color gradient. The image pre-processing, threshold segmentation, erosion and dilation operations based on satellite imagery road edge detection realize a rapid extraction method. II. THE BASIC PRINCIPLES AND PROCESS OF ROAD EXTRACTION The road feature extraction is divided into four main steps: 1. Road features enhancement: use image processing techniques, such as image wavelet transform, and image sharpening, etc.;

Edge detection or template matching method on road detection operator to determine the possible road point; 3. Use edge points to automatically connect or the Kalman filter method may point the road extension segment; 4. Detect and confirm the road segment, and the automatic connection into the road. In accordance with the theory of Marr vision, the vision from the initial original data (2-dimensional image data) to the final expression of the threedimensional environment has experienced three levels of low, medium and high. The road information extracted from remote sensing images, which should be three levels from low to high. The extraction process is shown in Figure 1. Image pre-processing (correction, filtering, etc.)

2.

III. USM SHARPENING TO ACHIEVE ROAD EXTRACTION IN HIGH-RESOLUTION SATELLITE IMAGES A. USM Sharpening Principle The purpose of the image sharpening process is to enhance the outline of the image, enhance the portion of the color jump in the image, highlighting the image detail. Commonly referred to as gou bian technology, it is the sharpening of the image processing. General sharpening of the image must have higher signal-tonoise ratio, otherwise, with image sharpening, the signal-to-noise ratio may even be lower. The sharpening will make the noise stronger than signal enhancement, so it must be handled with care. The traditional method of sharpening is achieved by calculating the gradient of the field of the images image point (or differential). Common methods are the Robert operator, Sobel operator, Laplace transform and Gauss Laplace (LoG) transform, etc. The above methods are based on gray image arithmetic. Since the color image is converted to the gray images, and the gray image is converted to a binary image will cause the loss of useful information. Therefore, using USM sharpening of color images can be converted to gray (or binary) image prior in order to maximize the use of the color images useful information. USM sharpening is a commonly used technique, and is used to sharpen the edge of the image. Use of the algorithm can quickly adjust the contrast of the image edge detail, generate a bright line a dark line on both sides of the edges of the image, and make the overall picture more clear. For high-resolution output, usually sharpening effect is displayed on the screen and is sharper than in printing. The principle is low-pass filtering (such as mean filter, median filter, and Gaussian filter) to give a smoother blurred image in the original image, and can enhance the color jump section strength. USM sharpening of the image is divided into a combination of high-frequency and low-frequency, and high-frequency and low-frequency signal processing, high-frequency signal enhancement expressions (1): (1) A* = A + f ( A B)

The low-level processing (image enhancement, segmentation, edge detection, texture and shape feature extraction)

Medium-level processing (analysis, select, comprehensive, description)

High-level processing (knowledge, rules, learning, understanding, reasoning, recognition, etc.)

Road (identification, show, application, etc.)


Figure1. Course of road feature extraction

For the extraction of the roads, first analyze the imaging features of the road, i.e. the image understanding, to understand the process, including artificial intelligence, computer vision, pattern recognition, mathematical modeling, and image processing, etc. In high resolution remote sensing images the obvious features of road, with the improvement of image resolution, image detail feature extraction technology development, more and more road goals, followed by the noise of the non-target image is also increasingly greater. This paper proposes the USM method to achieve the pre-processing of images to enhance the road features, in order to achieve the automatic extraction of road on the high-resolution remote sensing images.

B is A low-pass filtering the image, and f is the f is generally weighting coefficient, f [1, ] , but

A* is the enhanced image, A is the original image,

limited to <10, by adjusting the weighting coefficient, which can control the enhanced strength. The low-frequency signal suppression expression (2):

A = A + F ( B)

(2)

where A is the image after change, A is the original image, B is A low-pass filtered image, and F is a weighting coefficient, F [0,1]

From Formulas (1) and (2), it can be concluded that this methods basic algorithm is as follows (3): (3) g ( x, y ) = f ( x, y ) + C[ f ( x, y ) h( x, y )] wherein the for the artificial method Fuzzy f ( x, y ) of the obtained image; C is a constant, and for controlling the degree of sharpening of the image. In the second term of Formula (3), eliminating the majority of the lowfrequency portion of the original image, to retain the high frequency portion, the constant C is superimposed. This difference is several times the highfrequency component of the enhanced image. In achieving, h( x, y ) available simple local average method, field size to take a 3 3 window, mask for:

Originality image Finally image Filter image

f ( x, y )

is the original image,

h ( x, y )

For each point calculated in the image h( x, y ) , after substituting into Formula (3) that can be obtained by g ( x, y ) value. For example, when taken corresponding to the mask of Formula (3) as:

1 1 1 1 W = 1 1 1 9 1 1 1

(4)

Figure 2. USM sharpening principle

C =9,

B. Rapid Extraction Achieves High-Resolution Satellite Imagery Road in Matlab 1) The USM Sharpening Calculates High-Resolution Satellite Images: USM sharpening is used in order to emphasize the edge and produces a sharper image. For high-resolution satellite image roads, the roads are generally mostly dark cement, sometimes there is some dirt road, and with the surrounding environment color similar to the color discrimination becomes very difficult. It makes the almost similar dirt road, through the use of USM sharpening becomes a visible edge of the line between their environment and enhances the color gradient intensity between the road and surrounding environment, and the road edge is easy to extract. For color images, the use of the USM sharpening has a better effect than several other methods sharpening. Figure 3 is a high-resolution color satellite image comparison chart of different methods of sharpening results. (a) Original image; (b) USM sharpening the image processed; (c) Robert cross gradient processing image;, (d) Sobel gradient processing image.

It can calculate a different mask according to the value of C. When applied to each point in the image, you can quickly and easily achieve image sharpening. However, this method enhances the high-frequency component of the image as a whole, while highlighting the edges and contours of the amplified high-frequency noise, resulting in noticeable glitch noise [9]. Using a different low-pass filtering method for the final sharpening effect is not the same, compared to the other low-pass filter, and using the Gaussian blur can play the role of suppressing noise. Using the Gaussian low-pass filter of the lowest degree of blur of the image, the edge brought minimal aliasing, and the edges of the image will not be too sharp. Therefore, use USM sharpening Gaussian low-pass filter for filtering. Gaussian function-shaped, such as formula (6):
( x2 + y 2 )

1 1 1 w1 = 1 9 1 1 1 1

(5)

g ( x, y ) = e

2 2

wherein the is used to control the degree of smoothing; when is small, the filter positioning accuracy is high, but the SNR is low; when is large, the situation is just the opposite.

(6)

x R, y R

is

the size of the area of influence. The sharpening principle is shown in Figure 2.

(a) Original image

(b) r=3, n=2

(c) r=7, n=4 (a) Original image (b) USM sharpening image processed the

(d) r=9, n=3

Figure 4. Different form radius and processing times results contrast diagram

Because the USM is to color image RGB, three components adopt the filtering method, and the sharpen will cause color drift, making the environment color and color transition to red, green, blue, white, and black color transformation. So the processing times is unfavorable and overmuch, generally not more than 4 times.
(c) Robert cross gradient processing image; (d) Sobel gradient processing image Figure 3. Several different sharpening method processing results after contrast image

The image contrast can be seen, and the USM sharpening processing later image is closer to the original image, but also for the linear graphical representation is clearer than the original image. The USM sharpening method realization in Matlab color image processing algorithms as follows: I=imread(*.jpg); H=fspecial(gaussian,r,s); % r as template size, s is a smoothing parameter J=imfilter(I,H); I=2*I-J; For the different colors of the color image, we can use a different size template with single or multiple sharpening processing, until we achieve an ideal effect. See Figure 4, wherein r as template radius, n is the processing times.

2) Image Segmentation Operation: Image segmentation occupies a very important position in image analysis, image recognition, and image detection. Image segmentation is where the target of interest is extracted from the image; a specific area and the other portion are separated and extracted from the processing. In the image extraction, it is generally in accordance with the gray value of each pixel in the image and is greater than or less than the threshold value Th for image segmentation. The mathematical model of the threshold method of the following formula (7): Assume that the original image for the processed image segmentation

g ( x, y)

f ( x, y ) , after for g ( x, y ) ,
(7)

for binary images, it is:

The formula indicates that the threshold core is the method of determining the threshold Th . IV. ANALYSIS OF EXPERIMENTAL RESULTS General road are extracted in two ways: one is the middle line extraction, and the other is a contour extraction. The experiment used the latter extraction method. The image in the high-resolution color satellite image is used to complete the pretreatment on the basis of threshold segmentation operation, corrosion and

1, f ( x, y ) Th g ( x, y ) = 0, f ( x, y ) Th

expansion operations, and finally, using the edge detection algorithm to achieve the extraction of the road contour. We selected experimental data, Pictured Quickbird panchromatic image (spatial resolution 0.7m, size is 256 256). Through experiment calculation im2bw ( ) function in matlab segmentation calculation, and the threshold value is 55. Next, on the segmentation result of the operation, the expansion calculation, and finally using the Canny edge detection operator to perform contour extraction, and the experimental results for Figure 5 are below:

REFERENCES
[1] Z. G. Wang, Y. C. Wan, and H. Mao. A new method of extracting road from high-resolution remote sensing image, Geomatics & Spatial Information Technology, no. 6, pp. 149151, 2010. M. Wang, X. M. Yang, and G. N. Lv. A semi-automatic road extracting method for high spatial resolution remotely sensed imagery, Journal of Image and Graphic, no. 7, pp. 1322-1328, 2008. Q. Ye, X. H. Zhzng, and D. Wang. A road extraction method from high-resolution remote sensing image based on the regional growth and space shape constraints, Theory Research, no. 2, 2010.

[2]

[3]

(a) Original image

(b) USM algorithm to extract road image

(c) No USM algorithm extracts road image Figure 5. The USM algorithm and not the USM algorithm road extraction result contrast diagram

The USM algorithm can enhance the edges and details of the color image, while suppressing noise. By sharpening the image, extended image information is useful to highlight the contrast of the image, making image edges and details become clear, and it is quick and easy to extract roads from high-resolution satellite images. Facilitating the programming of the algorithm can handle full-color images, and the method is also suitable for the extraction of full-color topographic map contour lines and other linear images. However, this method also has its shortcomings: the maximum and minimum pixel value will exceed the original image sharpening, thus resulting in a change of image color difference, and then the choice of the template is very important. An improper selection can cause excessive enhancing or enhancing enough to affect the final extraction.

[4]

[5]

[6]

[7] [8] [9] [10]

[11] [12]

J. Xia, M. L. Hu, and H. L. Bu. Algorithm of road extraction from high resolution remote sensing images based on HopAlong, Journal of Hefei University of Technology, no. 12, pp. 19938-1940, 2009. Y. Fu, Z. Y. Qin, and J. J. Gong. Research actuality and thoughts on the automatic extraction of linear objects from remote sensing images, Journal of Geodesy and Geodynamics, no. 5, 2002. P. F. Wang, L. Wang, X. Z. Feng, and P. F. Xiao. Review of road extraction from remote sensing image, Remote Sensing Technology and Application, no. 6, 2009. Y. P. Hu, and L. T. Zhang. Advances in road feature extraction, Geospatial Information, no. 6, 2010. R. J. Offen. VLSI image processing (translation). Beijing: Science Press, 1989. (in Chinese). L. H. Zou. A new method of sharpening blurred figure images, Microcomputer Applications, no. 7, 2011. F. Liu, D. S. Liu, and Z. G. Gui. Image sharpening method based on Markov random field, Computer Engineering, no. 1, 2011. T. Cai, and R. S. Wang. An algorithm for extracting road network from multi-band remote sensing images, Journal of Software, vol. 12, no. 6, pp. 943-948, 2001. A. K. Shackelford, and C. H. Davis. Fully automated road network extraction from high-resolution Satellite multispectral imagery, in International Geoscience and Remote Sensing Symposium, Toulouse, France, 2003, p. 7.

[13] [14]

X. Y. Hu. Automatic extraction of linear features and buildings from airborne imagery. Doctoral dissertation, Wuhan University, 2001, p. 11. G. Vosselman, and J. de Knecht. Road tracing by profile matching and Kalman filtering, in Automatic Extraction of Man-made Objects from Aerial and Space Images, Monte Verrita, Birkhauser Verlag Basel, 1995, pp. 265-274.